[Feature Request]: Support for autoresearch as an Agent Tool #590

New issue

Open

opened 2026-06-01 17:18:15 +02:00 by Giver- · 0 comments

Giver- commented

2026-06-01 17:18:15 +02:00

(Migrated from github.com)

The Idea
I would love to see autoresearch integrated as a native "Skill" or "Tool" within Odysseus. This would allow an agent to autonomously perform ML research experiments and iterations directly from the workspace.

Why it's useful
Currently, autoresearch is a standalone script that requires a separate environment. Bringing this into Odysseus would allow users to:

Trigger an "Auto-Research" session via the chat interface.
Have the agent monitor the training loop, Git history, and system resources.
Report back on the research progress in the chat once the experiments are done.

How it could work

Define a new "Research Agent" tool in the Odysseus ecosystem.
The tool would need the ability to execute the autoresearch loop (which requires git, uv, and GPU access) in a scoped environment, perhaps handled via Docker containerization or specific user-defined execution paths on the host machine.
The agent could then parse the results.tsv or program.md files generated by autoresearch to provide summaries back to the user.

Context
Reference: https://github.com/karpathy/autoresearch

**The Idea** I would love to see `autoresearch` integrated as a native "Skill" or "Tool" within Odysseus. This would allow an agent to autonomously perform ML research experiments and iterations directly from the workspace. **Why it's useful** Currently, `autoresearch` is a standalone script that requires a separate environment. Bringing this into Odysseus would allow users to: 1. Trigger an "Auto-Research" session via the chat interface. 2. Have the agent monitor the training loop, Git history, and system resources. 3. Report back on the research progress in the chat once the experiments are done. **How it could work** - Define a new "Research Agent" tool in the Odysseus ecosystem. - The tool would need the ability to execute the `autoresearch` loop (which requires `git`, `uv`, and GPU access) in a scoped environment, perhaps handled via Docker containerization or specific user-defined execution paths on the host machine. - The agent could then parse the `results.tsv` or `program.md` files generated by `autoresearch` to provide summaries back to the user. **Context** Reference: https://github.com/karpathy/autoresearch